io_u: fall through to unlock path in case of error

[fio.git] / engines / libaio.c

diff --git a/engines/libaio.c b/engines/libaio.c
index f7e2f8967c8900a1d5c6839dda1d746df5986b60..b241ed99c4972e9956597d355645c65fa8274498 100644 (file)
--- a/engines/libaio.c
+++ b/engines/libaio.c
@@ -4,14 +4,17 @@
  * IO engine using the Linux native aio interface.
  *
  */
-#include <stdio.h>
 #include <stdlib.h>
 #include <unistd.h>
 #include <errno.h>
-#include <assert.h>
 #include <libaio.h>
 
 #include "../fio.h"
+#include "../lib/pow2.h"
+#include "../optgroup.h"
+
+#define IOCB_CMD_PREAD_POLL 9
+#define IOCB_CMD_PWRITE_POLL 10
 
 static int fio_libaio_commit(struct thread_data *td);
 
@@ -37,8 +40,9 @@ struct libaio_data {
 };
 
 struct libaio_options {
-       struct thread_data *td;
+       void *pad;
        unsigned int userspace_reap;
+       unsigned int hipri;
 };
 
 static struct fio_option options[] = {
@@ -51,6 +55,15 @@ static struct fio_option options[] = {
                .category = FIO_OPT_C_ENGINE,
                .group  = FIO_OPT_G_LIBAIO,
        },
+       {
+               .name   = "hipri",
+               .lname  = "RWF_HIPRI",
+               .type   = FIO_OPT_STR_SET,
+               .off1   = offsetof(struct libaio_options, hipri),
+               .help   = "Use polled IO completions",
+               .category = FIO_OPT_C_ENGINE,
+               .group  = FIO_OPT_G_LIBAIO,
+       },
        {
                .name   = NULL,
        },
@@ -68,12 +81,17 @@ static inline void ring_inc(struct libaio_data *ld, unsigned int *val,
 static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u)
 {
        struct fio_file *f = io_u->file;
+       struct libaio_options *o = td->eo;
 
-       if (io_u->ddir == DDIR_READ)
+       if (io_u->ddir == DDIR_READ) {
                io_prep_pread(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
-       else if (io_u->ddir == DDIR_WRITE)
+               if (o->hipri)
+                       io_u->iocb.aio_lio_opcode = IOCB_CMD_PREAD_POLL;
+       } else if (io_u->ddir == DDIR_WRITE) {
                io_prep_pwrite(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
-       else if (ddir_sync(io_u->ddir))
+               if (o->hipri)
+                       io_u->iocb.aio_lio_opcode = IOCB_CMD_PWRITE_POLL;
+       } else if (ddir_sync(io_u->ddir))
                io_prep_fsync(&io_u->iocb, f->fd);
 
        return 0;
@@ -81,7 +99,7 @@ static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u)
 
 static struct io_u *fio_libaio_event(struct thread_data *td, int event)
 {
-       struct libaio_data *ld = td->io_ops->data;
+       struct libaio_data *ld = td->io_ops_data;
        struct io_event *ev;
        struct io_u *io_u;
 
@@ -143,9 +161,9 @@ static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
 static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
                                unsigned int max, const struct timespec *t)
 {
-       struct libaio_data *ld = td->io_ops->data;
+       struct libaio_data *ld = td->io_ops_data;
        struct libaio_options *o = td->eo;
-       unsigned actual_min = td->o.iodepth_batch_complete == 0 ? 0 : min;
+       unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
        struct timespec __lt, *lt = NULL;
        int r, events = 0;
 
@@ -169,7 +187,8 @@ static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
                        events += r;
                else if ((min && r == 0) || r == -EAGAIN) {
                        fio_libaio_commit(td);
-                       usleep(100);
+                       if (actual_min)
+                               usleep(10);
                } else if (r != -EINTR)
                        break;
        } while (events < min);
@@ -177,9 +196,10 @@ static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
        return r < 0 ? r : events;
 }
 
-static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u)
+static enum fio_q_status fio_libaio_queue(struct thread_data *td,
+                                         struct io_u *io_u)
 {
-       struct libaio_data *ld = td->io_ops->data;
+       struct libaio_data *ld = td->io_ops_data;
 
        fio_ro_check(td, io_u);
 
@@ -205,6 +225,8 @@ static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u)
                        return FIO_Q_BUSY;
 
                do_io_u_trim(td, io_u);
+               io_u_mark_submit(td, 1);
+               io_u_mark_complete(td, 1);
                return FIO_Q_COMPLETED;
        }
 
@@ -218,7 +240,7 @@ static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u)
 static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
                              unsigned int nr)
 {
-       struct timeval now;
+       struct timespec now;
        unsigned int i;
 
        if (!fio_fill_issue_time(td))
@@ -236,10 +258,10 @@ static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
 
 static int fio_libaio_commit(struct thread_data *td)
 {
-       struct libaio_data *ld = td->io_ops->data;
+       struct libaio_data *ld = td->io_ops_data;
        struct iocb **iocbs;
        struct io_u **io_us;
-       struct timeval tv;
+       struct timespec ts;
        int ret, wait_start = 0;
 
        if (!ld->queued)
@@ -280,9 +302,9 @@ static int fio_libaio_commit(struct thread_data *td)
                                break;
                        }
                        if (!wait_start) {
-                               fio_gettime(&tv, NULL);
+                               fio_gettime(&ts, NULL);
                                wait_start = 1;
-                       } else if (mtime_since_now(&tv) > 30000) {
+                       } else if (mtime_since_now(&ts) > 30000) {
                                log_err("fio: aio appears to be stalled, giving up\n");
                                break;
                        }
@@ -306,17 +328,24 @@ static int fio_libaio_commit(struct thread_data *td)
 
 static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u)
 {
-       struct libaio_data *ld = td->io_ops->data;
+       struct libaio_data *ld = td->io_ops_data;
 
        return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
 }
 
 static void fio_libaio_cleanup(struct thread_data *td)
 {
-       struct libaio_data *ld = td->io_ops->data;
+       struct libaio_data *ld = td->io_ops_data;
 
        if (ld) {
-               io_destroy(ld->aio_ctx);
+               /*
+                * Work-around to avoid huge RCU stalls at exit time. If we
+                * don't do this here, then it'll be torn down by exit_aio().
+                * But for that case we can parallellize the freeing, thus
+                * speeding it up a lot.
+                */
+               if (!(td->flags & TD_F_CHILD))
+                       io_destroy(ld->aio_ctx);
                free(ld->aio_events);
                free(ld->iocbs);
                free(ld->io_us);
@@ -354,7 +383,7 @@ static int fio_libaio_init(struct thread_data *td)
        ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
        ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
 
-       td->io_ops->data = ld;
+       td->io_ops_data = ld;
        return 0;
 }